Conversion measures how often an eligible person or event completes a specified transition. In image search, conversion rate (CR) is the share of clicks that lead to a purchase. State both the action and its denominator: purchases per click answers a different question from purchases per offer shown. section 5.3.2 section 5.2.2
Choose the transition to match the decision the system is meant to improve. A high conversion rate is not automatically a successful business outcome, and a promising validation result does not by itself establish production impact. section 5.2 section 4.2
When using conversion to guide an ML system:
- Specify the eligible population and completed action. For image search, track clicks that become purchases alongside click-through rate, which tracks clicks on results. section 5.3.2
- Place conversion in a metric hierarchy. An e-commerce recommender may seek more purchases to increase gross merchandise value (GMV), but that link is an assumption to test. section 5.2.2
- Plan for sparse outcomes. If 3% of offers receive clicks and 3% of those clicks lead to purchases, only 9 of 10,000 offers produce a purchase. Click-through rate may be a faster proxy, provided its relationship with conversion is checked. section 5.2.2
A bank’s payment transition
A bank tracked clients moving from nonpayer to payer. Its existing conversion rate was 0.5, amid calls that could secure promises without actual payment. A system predicting agreement and fulfillment, paired with text messages instead of calls, reached 0.9 conversion in validation. The relevant transition was payment, not merely a promise to pay. section 5.2
A better number can still miss the question
Another collections model selected whom to contact and tested at roughly 80% conversion, against about 50% for the manual process. Leadership also wanted to understand customers’ motivations. Conversion showed whom the model could activate; it did not answer why. section 4.2
Treat conversion as an online outcome when it depends on live behavior. For image search, CR and click-through rate require a live system; ranking measures such as mean reciprocal rank and normalized discounted cumulative gain can serve as offline proxies. Check that proxy improvements translate into the conversion and business outcomes that matter. section 5.3.2 section 5.2.2
